Learning path
Fabric: InfiniBand and Spectrum-X Ethernet
How GPUs talk to each other across racks. Two fabrics compete: InfiniBand, and NVIDIA Spectrum-X, which brings lossless Ethernet (RoCE), adaptive routing and telemetry to standard Ethernet. The exam weighs them equally at 30 % each; troubleshooting tools are another 20 %.
Source: https://neocloud.fyi/sources/ Checked 2026-09-18
site's own compiled learning path, synthesised from the sources cited in each step
- optional Introduction to Networking Generic networking primer. Skim if you already run networks. Opened 2026-10-06: NVIDIA's AI networking learning path lists this course as self-paced, free, 2 hours: https://www.nvidia.com/en-us/learn/learning-paths/ai-networking/
- core InfiniBand Essentials Short official primer: subnet manager, LIDs, partition keys, the vocabulary of the fabric.
- core Cumulus Linux Essentials The Cumulus course NVIDIA actually links from the exam page: CLI, interfaces, routing, bonds, bridges.
- core InfiniBand Network Administration The one paid course worth it: configuration, QoS, adaptive routing, partition keys, UFM. Maps onto a full 30 % domain.
- core NVIDIA DSX Air simulator Cloud simulator for Cumulus switches, Spectrum-X, SuperNICs and NetQ. Rehearse NVUE, BGP-EVPN and RoCE configuration without hardware. Confirm the free tier at sign-up. Opened 2026-10-06: NVIDIA's DSX Air account-setup docs describe a one-year free trial (60 vCPUs, 60 GiB, 10,000 compute hours) started by an NGC organisation owner; a business email is required and usage is otherwise billed per compute hour: https://docs.nvidia.com/dsx-air/account-setup/
-
core
Read the docs instead of a course
UFM overview and diagnostics, the NVUE reference, the NCCL guide and the MLNX_OFED to DOCA-OFED transition guide.
Also: NVUE reference · NCCL guide · DOCA-OFED transition - optional Introduction to DOCA for DPUs Free two-hour course on BlueField DPUs and the DOCA framework, the part of the fabric that lives in the server.
-
core
Half a day on a rented InfiniBand cluster
Two GPU nodes with real InfiniBand: run nccl-tests, perftest, ibstat and ibdiagnet, and keep the numbers. Verda and RunPod bill by the hour with no minimum; Lambda needs 16 GPUs for two weeks.
Also: RunPod Instant Clusters · nccl-tests - skip Cumulus Linux Administration Deeper paid boot camp. The free Essentials course plus the simulator covers what the exam asks. Opened 2026-10-06: NVIDIA's AI networking learning path lists this course as self-paced, USD 100, 5 hours: https://www.nvidia.com/en-us/learn/learning-paths/ai-networking/
- skip Spectrum-X Administration, instructor-led Twelve hours for USD 1,500. The documentation and simulator cover the same objectives. Opened 2026-10-06: NVIDIA's AI networking learning path lists the Spectrum-X Networking Platform Administration private workshop as instructor-led, USD 1,500, 12 hours: https://www.nvidia.com/en-us/learn/learning-paths/ai-networking/
- core Exam: NCP-AIN See the exam entry below.
What was not found
No first-person pass reports, no official practice exam and no credible video series were found. Sites selling 'actual questions' are dumps and break the exam agreement.
Connects to
- Cisco 300-640 DCAI, Data Center AI Infrastructure
- Cooling: liquid to the chip, heat to the air Hoses and NVLink copper share the rear of the rack; switch racks stay air-cooled.
- Facility: medium voltage to the rack Cable limits set row pitch and tray sizes. Every 400G or 800G optic draws roughly 10 to 30 W, which lands in the rack power and cooling budget.
- GPUDirect Storage; VAST, WEKA, DDN
- InfiniBand vs Spectrum-X
- NCCL tests and burn-in
- NCP-AIN, AI Networking Professional
- Operations: Slurm, Kubernetes and fleet management The Slurm topology plugin and NCCL both model this fabric. The NVIDIA Network Operator (5 % of the exam) is the seam with Kubernetes.
- Racks: what an NVL72-class rack needs Passive copper reaches 1.5 to 3 m, active copper about 7 m, optics 50 m to 2 km. Those limits decide which racks must stand together.
- Structure and layout
- UFM, NetQ, WJH